dsh-plugin-usage

dsh-plugin-usage

Deployment-level 用量 / Usage plugin for DeepSeek Harness (DSH).

A real-time token & cost dashboard tab for the current session, placed next to Chat / Trajectory in the conversation view ring. It records input / output / cache-hit tokens and prices every model call with both the DeepSeek official peak/off-peak rate card and the OpenCode Go plan rate card, refreshes every second, shows animated gradient-blue status bars, supports a fully editable custom price table (persisted across restarts) and CSV/JSON export.

中文文档见 README.zh.md（Chinese documentation is in README.zh.md）。

---

✨ Features

• New "用量" tab in the conversation view ring — sits right after Trajectory (order: 11), native DSH UI style (theme tokens, compact toolbar, card sections).
• Real-time live tracking — wraps every llm/stream model call: a "generating" row appears the moment a call starts, output/reasoning tokens are estimated per chunk and updated every second, then replaced by the adapter's exact usage; the committed assistant/message finalizes the record (15s match window, no duplicates).
• Per-session accounting — input (cache miss), output (incl. reasoning), cache hit, cache write; totals and per-call details with timestamps.
• History preload — the first time a session is opened, the full history is rebuilt from the persisted session log (seq-level dedup against live collection).
• Dual pricing, auto-adjusted — every call is priced by its own timestamp:
  • DeepSeek official (CNY / 1M tokens, effective 2026-08-17): peak hours (Beijing 09:00–12:00, 14:00–18:00) are 2× off-peak. V4 Pro: peak 9/0.3/27, off 4.5/0.15/13.5 (in/cache/out); V4 Flash: peak 3/0.1/9, off 1.5/0.05/4.5.
  • OpenCode Go plan (USD / 1M tokens + monthly quota): 20 built-in models (DeepSeek V4 Pro/Flash, Grok 4.5, GPT 5.6 Luna, GLM, Kimi, MiMo, MiniMax, Qwen3, Hy3), e.g. V4 Flash off-peak 0.22/0.007/0.66 and peak 0.44/0.014/1.32 with $15/month quota (DeepSeek plan models are billed in two peak tiers); plan is $10/month ≈ 6× usage value.
  • Three display modes: Auto (both), Official only, Plan only.
• Animated gradient-blue status bars — input / output / cache-hit distribution, smooth width transitions refreshed every second.
• OpenCode Go official quota — reads the official account quota (rolling 5h / weekly / monthly windows with reset times) straight from the opencode.ai/zen/go/v1/usage API, matching the console numbers.
• Custom price table — add, edit or remove prices for any model (official peak/off + plan prices + quota). Custom entries override the built-in cards, are persisted to <workspaceRoot>/.dsh-usage-prices.json, and historical costs are re-priced immediately after a change.
• CSV / JSON export — one-click download links (60s validity) for the full call detail list with per-call prices and totals.
• Peak/off-peak indicator — the toolbar shows whether the current Beijing time is peak or off-peak.

🔐 Privacy

• No API keys, credentials, tokens or machine-specific paths are read, logged, or stored. Only usage numbers are kept in memory.
• The custom price file stores prices only (no conversation content, no tokens, no keys).
• The plugin never sends data anywhere: all computation is local; exports are served from the local DSH HTTP server as one-shot links.

🚀 Install

⚠️ Important: register the plugin exactly once. Do not combine installation paths.

• dsh plugin add (Method A / B) adds the plugin to dsh.profile.bundles; the bundled cordis.patch.yml is applied automatically as a bundle layer — do not also add ui-usage manually.
• ./install.sh or manual patching (Method C / D) registers through cordis.patch.yml — do not also use dsh plugin add.
• If two ui-usage entries exist, DSH fails with duplicate loader entry id: ui-usage. Fix: remove one. ./install.sh automatically removes the plugin from dsh.profile.bundles to prevent this.

Prerequisites

• DeepSeek Harness installed and the web profile initialized at least once (dsh web), so ~/.dsh/profiles/web exists.

Method A: install from the GitHub repository (recommended)

dsh plugin --profile web add github:GHJIVHIDD/dsh-plugin-usage

Pin a commit for reproducibility:

dsh plugin --profile web add github:GHJIVHIDD/dsh-plugin-usage#<commit-sha>

pnpm ≥ 10 build authorization: if the first run fails because pnpm refuses to run the git dependency's prepare script, add this to the profile's pnpm-workspace.yaml:

# File: ~/.dsh/profiles/web/pnpm-workspace.yaml
allowBuilds:
  "@deepseek-ai/dsh-plugin-usage": true

Then re-run the install command.

Method B: install from the GitHub Release tarball

The install package is published in GitHub Releases (not in the source tree):

curl -L -o dsh-plugin-usage-0.1.5.tgz \
  https://github.com/GHJIVHIDD/dsh-plugin-usage/releases/download/v0.1.5/dsh-plugin-usage-0.1.5.tgz

dsh plugin --profile web add ./dsh-plugin-usage-0.1.5.tgz

Or download it from the Releases page:

https://github.com/GHJIVHIDD/dsh-plugin-usage/releases

You can also add a local source directory:

dsh plugin --profile web add /path/to/dsh-plugin-usage

Method C: no-pnpm manual install script

git clone https://github.com/GHJIVHIDD/dsh-plugin-usage.git
cd dsh-plugin-usage
./install.sh

The script copies the plugin into ~/.dsh/profiles/web/node_modules/@deepseek-ai/dsh-plugin-usage, writes the ui-usage patch into cordis.patch.yml, and removes the plugin from dsh.profile.bundles (with a package.json.bak backup) if present. Customize the target:

DSH_HOME=/path/to/.dsh DSH_PROFILE=web ./install.sh

Method D: manual copy + patch

# 1. Copy the package into the profile's node_modules
mkdir -p ~/.dsh/profiles/web/node_modules/@deepseek-ai
cp -R dsh-plugin-usage ~/.dsh/profiles/web/node_modules/@deepseek-ai/

# 2. Make sure ~/.dsh/profiles/web/cordis.patch.yml contains:
# - insert:
#     - id: ui-usage
#       name: '@deepseek-ai/dsh-plugin-usage'

Verify the installation

# The profile must compose without errors:
dsh --profile web --dump-config > /dev/null && echo OK

Then start (or refresh) your web profile and open the 用量 tab:

dsh web

If the tab is missing, check the browser console for usage-api fetch failures, and confirm the patch entry exists exactly once (grep -n ui-usage ~/.dsh/profiles/web/cordis.patch.yml ~/.dsh/profiles/web/package.json).

🧭 Usage guide

1. Start a conversation and let the agent run (each model call is tracked).
2. Open the 用量 tab (next to Trajectory).
3. The page refreshes every second:
   • Session overview — input / output (incl. reasoning) / cache hit / cost (official ¥ and plan $), current model and call count.
   • Usage distribution — gradient-blue bars for input, output and cache hit (with a pulsing dot and "generating" label on the in-flight call).
   • OpenCode Go official quota — rolling (5h) / weekly / monthly usage percentages with reset times, identical to the OpenCode console; requires the API key (see below).
   • Call details — time, model, input/output/cache tokens and per-call official ¥ / plan $ costs; the in-flight row shows a blue pulse indicator.
   • Price table (collapsible) — full merged rate card; Edit any model, + Add model for new ones, Remove for custom entries.
4. Export — click Export in the Call details header, then download CSV or JSON (links valid for 60 s; click again to regenerate).
5. Pricing mode — toolbar switches: Auto (both) / Official / Plan.

OpenCode Go official quota

• The official quota card queries GET https://opencode.ai/zen/go/v1/usage with your OpenCode Go API key and shows the account-level rolling (5h) / weekly / monthly usage percentages and reset times — the same numbers as the OpenCode console.
• The key is read from the environment only (never stored or logged), in order:
  • DSH_OPENCODE_GO_KEY — dedicated key (recommended), or
  • OPENCODE_GO_API_KEY — the env name used by the llm-pi-ai opencode-go provider config, or
  • DEEPSEEK_API_KEY — used automatically when your deepseek provider points at the OpenCode Go gateway.
• Set it before starting dsh, e.g. export DSH_OPENCODE_GO_KEY=sk-..., then restart. The card refreshes every 30 s.
• The session-scoped plan cost (¥/$ in the overview) is separate from the account quota: it reflects only this session.

Custom price table

• Units: official CNY per 1M tokens (peak & off-peak), plan USD per 1M tokens + monthly quota.
• Leave a field empty to omit that price group field.
• Changes apply immediately to the whole session history (dynamic re-pricing) and are saved to <workspaceRoot>/.dsh-usage-prices.json.
• Removing a custom entry falls back to the built-in card.

Peak / off-peak

Peak hours are Beijing time 09:00–12:00 and 14:00–18:00; official prices are 2× during peak. The toolbar badge shows the current period; per-call prices use the call's own timestamp, so history stays correct after the period changes.

🗂️ Package layout

dsh-plugin-usage/
├── package.json          # npm package manifest with dsh.bundle.patch + dsh.client metadata
├── cordis.patch.yml      # loader patch: inserts the ui-usage entry
├── install.sh            # no-pnpm manual installer (Method C)
├── lib/
│   ├── index.js          # host half (ESM): collection, pricing, /usage-api/* routes
│   ├── client.js         # client half (web ModuleLoader bundle)
│   └── types/index.d.ts  # TypeScript declarations
├── scripts/verify.mjs    # structural + privacy verification (dev-time)
├── README.md             # this file (English)
├── README.zh.md          # Chinese documentation
└── LICENSE               # MIT

HTTP API (internal, same-origin)

🧩 Compatibility

• Built and tested against DSH web profile (conversation view ring, conversation.view slot).
• Uses the same packaging conventions as dsh-plugin-canvas / dsh-plugin-vm-sandbox (dsh.bundle.patch + dsh.client metadata, ./client export, loader patch).
• No external runtime dependencies; the host half uses only DSH core services (webServer, timer, fs, sandboxPolicy, sessionQuery, llm/stream, session/event).

❓ FAQ

Q: The tab shows "Failed to load usage data". A: The host half is not serving /usage-api/*. Confirm the plugin is registered exactly once (see Install), then restart the profile. Check dsh --profile web --dump-config output for the ui-usage row.

Q: Why do numbers change after I edit a price? A: Costs are computed dynamically from the current price table on every read, so historical calls are re-priced immediately. This is by design.

Q: Is my conversation content stored? A: No. Only token counts, model names and timestamps are kept in memory; exports contain the same data. The custom price file contains prices only.

Q: Can I use it with the headless profile? A: The UI is web-only; the host half works in any profile, but the dashboard requires the web client.

📄 License

MIT © 2026 GHJIVHIDD